Datasets:

chrockey
/

DexJoCo-preprocessed

@@ -10,163 +10,86 @@ tags:
 # DexJoCo, preprocessed for PointWAM
-These are the per-demo frames that PointWAM is trained on for the [DexJoCo](https://github.com/brave-eai/dexjoco) benchmark
-(Franka + Allegro in MuJoCo, 11 tasks). Each released DexJoCo demo is replayed in the simulator and rendered
-to RGB-D at 224 x 224 with per-pixel body ids, camera poses, robot state and the recorded actions. There is
-one H5 file per demo, in one folder per task.
-Private repo. The read-only access token is in the `Checkpoints` section of the PointWAM GitHub README
-(github.com/chrockey/PointWAM, private). Without access to that, ask chrockey (Chunghyun Park).
 ## Versions
-**This repo is the one source of the DexJoCo training data for every PointWAM author and every cluster.** Each version is a
-git tag here. A tag is never moved or overwritten; a change to the data is a new commit and a new tag, added to the
-table below. Copies on other clusters (Foundry on kakao / MLXP / AWS, NHN, lab servers) are downloads of a tag, never
-local rebuilds, and are checked against that tag's manifest before training. Train with `--revision <tag>`, and
-record the tag with the result.
-| version | revision | date | H5 files | size | change |
 |---|---|---|---|---|---|
-| `v1.0` | `110a68c5` | 2026-10-06 | 1,059 | 113.5 GB | first upload: 224 x 224, 11 tasks, all 100 `bimanual_unlock_ipad` demos (42 recovered); the 41 demos that fail plain replay are missing |
-| `v1.1` | `421daed6` | 2026-10-28 | 1,100 | 119.9 GB | adds the 41 recovered demos; v1.0's 1,059 files are byte-identical |
-| `v1.2` | `72e98046` | 2026-10-28 | 1,100 | 119.7 GB | trims the 5 demos whose recording is frozen at the start (see "How it was made"); the other 1,095 files are byte-identical |
-Which version each PointWAM model trained on (the exact file lists are in `corpora/`):
-| model | version | demos | demo list |
 |---|---|---|---|
-| PointWAM v1.0.0 (ICLR submission) | `v1.0` | 1,017 | `corpora/pointwam-v1.0.0.txt`: v1.0 without the 42 recovered `bimanual_unlock_ipad` demos |
-| PointWAM v1.1.0, v1.1.1 | `v1.1` | 1,099 | `corpora/pointwam-v1.1.0.txt`: v1.1 without `bimanual_hanoi/hanoi_demo_11_2026-03-13_15-46-49_120095.h5` |
-| PointWAM v1.1.2 (in testing) | `v1.2` | 1,100 | `corpora/pointwam-v1.1.2.txt`: all of v1.2 |
-**Checking a copy.** `versions/<tag>.sha256` lists the sha256 of every H5 in a tag (in `sha256sum -c` format). The
-manifest files were added after the data commits, so take them from `main`:
-```bash
-hf download chrockey/DexJoCo-preprocessed --repo-type dataset --revision v1.2 --local-dir DexJoCo-preprocessed
-hf download chrockey/DexJoCo-preprocessed versions/v1.2.sha256 --repo-type dataset --local-dir manifests
-cd DexJoCo-preprocessed && sha256sum -c --quiet ../manifests/versions/v1.2.sha256   # prints nothing when every file matches
-```
-One number per tag, the sha256 of its manifest file. It equals
-`find . -name '*.h5' -print0 | LC_ALL=C sort -z | xargs -0 sha256sum | sha256sum` run in the copy's root:
-| version | manifest sha256 |
 |---|---|
 | `v1.0` | `1dca8061623db35321e1e3f8df73991ed314e398fca9d771f71204cda24ce5a3` |
 | `v1.1` | `91c2468b9bd8a411a918914794736c4d294529e3d6d4ea89c92f13422985f7b7` |
 | `v1.2` | `6ff159cb75a86051a46a705c13a4d9e22f26a093cdeefbf269c359d524cd9db9` |
-**Copies on RLWRLD clusters (Foundry).** A Foundry copy of a tag is named `pointwam_dexjoco_224_<version>_<revision>`
-(e.g. `pointwam_dexjoco_224_v1.2_72e98046`) and carries `hf_repo`, `hf_revision` and the manifest sha256 above in its
-metadata. The 256 x 256 DexJoCo frames already in Foundry (`dexjoco_frames_256_11task_bbb764fc_20260905`,
-`pointwam_dexjoco_stage03_*`) are earlier, separate renders, not versions of this dataset; results trained on them
-are not comparable with the models above.
-## Layout
-```
-<task>/<task>_demo_<k>_<date>_<time>.h5     one file per demo (bimanual_unlock_ipad files are named ipad_unlock_demo_*)
-```
-| task | arms | demos | success | frames | cameras | action dim | GB | prompt |
-|---|---|---|---|---|---|---|---|---|
-| `bimanual_assembly` | 2 | 100 | 100 | 72,642 | back, handcam_rgb_left, handcam_rgb_right | 46 | 16.5 | Grasp the tray with the left hand and the peg with the right hand, then insert the peg into the hole. |
-| `bimanual_hanoi` | 2 | 100 | 100 | 109,105 | back, handcam_rgb_left, handcam_rgb_right | 46 | 26.3 | Execute the final two moves of the three-level Tower of Hanoi: move the medium disk from the middle peg to the right peg with the right hand, then move the small disk from the left peg to the right peg with the left hand. |
-| `bimanual_microwave_cook` | 2 | 100 | 100 | 84,790 | back, handcam_rgb_left, handcam_rgb_right | 46 | 20.6 | Open the microwave door, place the food inside the microwave, close the door, and press the start button. |
-| `bimanual_photograph` | 2 | 100 | 100 | 40,255 | back, handcam_rgb_left, handcam_rgb_right | 46 | 8.3 | Grasp the camera with the left hand, align it with the logo, and press the shutter button with the right hand. |
-| `bimanual_unlock_ipad` | 2 | 100 | 100 | 53,144 | back, handcam_rgb_left, handcam_rgb_right | 46 | 12.1 | Grasp the iPad and enter the password 123 to unlock the device. |
-| `click_mouse` | 1 | 100 | 100 | 39,982 | front, handcam_rgb | 23 | 5.4 | Move the mouse to the purple mouse pad and click the left mouse button. |
-| `fold_glasses` | 1 | 100 | 100 | 57,900 | front, handcam_rgb | 23 | 8.8 | Fold the glasses and place them into the case. |
-| `hammer_nail` | 1 | 100 | 100 | 25,034 | front, handcam_rgb | 23 | 4.3 | Use the hammer to drive the nail into the wooden board. |
-| `pick_bucket` | 1 | 100 | 100 | 46,910 | front, handcam_rgb | 23 | 7.3 | Place the boxed food into the bucket and then lift the bucket. |
-| `pinch_tongs` | 1 | 100 | 100 | 43,441 | front, handcam_rgb | 23 | 5.9 | Grasp the tongs and perform three consecutive open-close motions. |
-| `water_plant` | 1 | 100 | 100 | 27,753 | front, handcam_rgb | 23 | 4.3 | Grasp the watering can and apply water to the plant. |
-All 1,100 released demos are included, 100 per task, and every one reaches the task's success predicate on replay.
-83 of them do not succeed when replayed as released; they were recovered (see "How it was made").
-## H5 schema (per demo, T = `n_frames`, frames at 30 Hz)
-| key | shape / dtype | meaning |
-|---|---|---|
-| attrs `task`, `prompt`, `n_arm`, `n_frames`, `size`, `success`, `cameras` | | task name, language instruction, 1 or 2 arms, T, image size (224), replay success, camera names |
-| attrs `trim_start`, `trim_reason` | | only in the 5 trimmed demos (see "How it was made"): frame 0 of the file is frame `trim_start` of the replay |
-| `actions` | (≤T, 23·n_arm) float32 | recorded command per arm `[pos3, quat4, hand16]`: Franka end-effector (TCP) target position + orientation quaternion, then the 16 Allegro joint targets; arms interleaved (arm 0 block, then arm 1). The env and PointWAM use 22/arm with the quaternion converted to a rotation vector. `actions[i]` is the command that produced state `i`. In some demos it stops at the end of the recording, and the frames after it (the held tail) have no action row; use `min` over the per-frame arrays |
-| `proprio` | (T, 23·n_arm) float32 | robot state, BLOCKED (not interleaved): `[tcp pose 7·n_arm (pos3 + quat4 per arm) \| hand joints 16·n_arm]` |
-| `qpos` | (T, nq) float64 | full MuJoCo `qpos` (robot + objects) |
-| `body_names`, `is_robot`, `is_hand` | (B,) | MuJoCo bodies and which of them belong to the robot / the hand |
-| `xpos`, `xquat` | (T, B, 3) / (T, B, 4) float32 | world pose of every body |
-| `<cam>/rgb` | (T, 224, 224, 3) uint8 | RGB |
-| `<cam>/depth` | (T, 224, 224) float16 | metric depth (m) |
-| `<cam>/body_id` | (T, 224, 224) uint16 | per-pixel MuJoCo body id (index into `body_names`) |
-| `<cam>/intrinsic` | (3, 3) float32 | pinhole intrinsics at 224 |
-| `<cam>/extrinsic` | (T, 4, 4) float32 | camera pose per frame (hand cameras move) |
-Large arrays are gzip-compressed. Read them with `h5py`.
-## How it was made
-- Environment: `brave-eai/dexjoco` at `8d23b0f`, plus two patches (`dexjoco-eval-inference-flag-deadlock`,
-  `dexjoco-drain-action-queue-on-reset`). Demos: [`DexJoCo/DexJoCo-Datasets-Raw`](https://huggingface.co/datasets/DexJoCo/DexJoCo-Datasets-Raw).
-- `pointwam.preprocessing.dexjoco.replay_capture` replays each demo and holds the last action for `--extend-steps`
-  more steps, because the released demos stop the moment success fires (60 steps for the single-arm tasks, 90 for
-  hanoi, 150 for unlock_ipad, 0 for photograph). Then `pointwam.preprocessing.dexjoco.extract_frames` renders the
-  frames at 224. Model fields that an env changes during an episode (the nail's mocap pose, display / button
-  colours) are recorded per frame and restored when rendering.
-- `hammer_nail` comes from the `mocaptail` capture: it renders the moving nail correctly and keeps 30 frames after success.
-- `bimanual_unlock_ipad`: 42 of the 100 released demos (all from the 2026-03-21 session) carry a wrong iPad pose and
-  never press a digit on replay (brave-eai/dexjoco#27). All 42 are recovered. For 39, the iPad xy is recovered by a
-  grid search over the iPad position. For the other 3 (`_13_10-27-33`, `_24_10-35-31`, `_33_10-40-33`), the stand/iPad
-  height is also restored to the env's own rule (0.92 m + table offset). The stand->iPad offset and the recorded
-  actions are unchanged, and success is the env's own `info["succeed"]`. With the 58 demos that replay as released,
-  that gives 100. The 42 recovered demos show the lit buttons and the unlocked screen; the 58 older captures do not.
-- The other 41 demos that fail on plain replay (pinch_tongs 14, bimanual_hanoi 14, fold_glasses 8, bimanual_assembly 4,
-  bimanual_photograph 1) are marginal contact outcomes: the release stores the initial object poses as float32, and
-  the replay flips with tiny numeric differences. 40 are recovered by a search over the initial object poses: 23 by
-  the plain replay, 17 with a position offset of at most 3.2 mm (search order: plain replay, then random offsets at
-  sigma 1 mm, then 3 mm; the smallest successful offset is kept; orientations and actions unchanged). `hanoi_demo_11_2026-03-13_15-46-49_120095` succeeds
-  on a plain replay on a different CPU (AMD EPYC 7452). All replays run with `OMP_NUM_THREADS=1`; without the
-  thread pin, marginal demos flip. Added 2026-10-28.
-- Every H5 was checked against its raw recording (`scripts/validate_dexjoco_frames.py` in the GitHub repo): actions
-  identical, replayed arm within 0.000 mm (median) of the recorded one, all structure checks pass. In 5 demos the raw
-  recording itself is broken at the start: the recorded arm state is frozen for 28-100 frames (after one glitch
-  frame) while the recorded commands already move the arm, so the replay leaves the recording there. These 5 are
-  TRIMMED to the frame where the replay first rejoins the recording within 10 mm (2026-10-28): click_mouse
-  `monitor_mousepad_demo_1_2026-03-10_20-03-12_829101` (first 143 frames dropped) and `_4_2026-03-10_20-05-08_469389`
-  (171); bimanual_microwave_cook `microwave_cook_demo_4_2026-03-10_18-45-41_789229` (376),
-  `_1_2026-03-10_17-39-11_295721` (177), `_1_2026-03-10_18-35-23_755710` (98). The untrimmed files, which PointWAM
-  v1.0.0-v1.1.1 were trained on, are in revision `421daed66ecb5f3aec2e3ea4faac41aac667784b` of this repo.
-## Download
 ```bash
-pip install -U huggingface_hub   # the `hf` CLI (huggingface-cli was removed in 1.x)
-hf auth login --token <read-only token, see above>
-hf download chrockey/DexJoCo-preprocessed --repo-type dataset --local-dir DexJoCo-preprocessed --include "water_plant/*"   # one task
-hf download chrockey/DexJoCo-preprocessed --repo-type dataset --local-dir DexJoCo-preprocessed                             # all (~120 GB)
 ```
-or from Python:
-```python
-from huggingface_hub import snapshot_download
-snapshot_download("chrockey/DexJoCo-preprocessed", repo_type="dataset", local_dir="DexJoCo-preprocessed",
-                  allow_patterns=["water_plant/*"])   # drop allow_patterns for all
-```
-Pin a version with `--revision <tag>` (CLI) or `revision="<tag>"` (Python); without it you get `main`, which can move.
-To train PointWAM on it: the GitHub README's `DexJoCo training data` section has the symlinks that put these
-folders where `configs/data/dexjoco/*.yaml` expect them. Which demos each released model trained on is in
-"Versions" above:
-| corpus | version | demos | config | model |
-|---|---|---|---|---|
-| everything | `v1.2` | 1,100 | `data=dexjoco/mt11_all1100` (branch `chunghyun/v112`) | PointWAM v1.1.2 (in testing) |
-| all but `hanoi_demo_11_2026-03-13_15-46-49_120095` | `v1.1` | 1,099 | `data=dexjoco/mt11_ipadrec_drec` | PointWAM v1.1.0 / v1.1.1 |
-| all but the 41 recovered demos | `v1.0` | 1,059 | `data=dexjoco/mt11_ipadrec` | — |
-| all but the 41 and the 42 recovered unlock_ipad demos | `v1.0` | 1,017 | `data=dexjoco/mt11` (`dexjoco_all11_mixed_224` in v1.0.0's code) | PointWAM v1.0.0 (ICLR) |
-The 41 are listed in `experiments/dexjoco-demo-recovery.md` of the GitHub repo.

 # DexJoCo, preprocessed for PointWAM
+PointWAM's training frames for [DexJoCo](https://github.com/brave-eai/dexjoco) (Franka + Allegro, MuJoCo, 11 tasks):
+one H5 per demo, RGB-D at 224 x 224 with body ids, camera poses, robot state and actions.
+Private; the read-only token is in the PointWAM GitHub README (`Checkpoints`), or ask chrockey.
 ## Versions
+This repo is the single source for every author and cluster. Each version is a tag; tags never move.
+Other clusters hold downloads of a tag, checked with its manifest.
+| tag | revision | date | demos | GB | change |
 |---|---|---|---|---|---|
+| `v1.0` | `110a68c5` | 10-06 | 1,059 | 113.5 | first upload (incl. 42 recovered unlock_ipad demos) |
+| `v1.1` | `421daed6` | 10-28 | 1,100 | 119.9 | + 41 recovered demos |
+| `v1.2` | `72e98046` | 10-28 | 1,100 | 119.7 | 5 frozen-start demos trimmed |
+| model | tag | demos | list |
 |---|---|---|---|
+| PointWAM v1.0.0 (ICLR) | `v1.0` | 1,017 | `corpora/pointwam-v1.0.0.txt` |
+| PointWAM v1.1.0 / v1.1.1 | `v1.1` | 1,099 | `corpora/pointwam-v1.1.0.txt` |
+| PointWAM v1.1.2 (testing) | `v1.2` | 1,100 | `corpora/pointwam-v1.1.2.txt` |
+Manifest sha256 (`find . -name '*.h5' -print0 | LC_ALL=C sort -z | xargs -0 sha256sum | sha256sum`):
+| tag | sha256 |
 |---|---|
 | `v1.0` | `1dca8061623db35321e1e3f8df73991ed314e398fca9d771f71204cda24ce5a3` |
 | `v1.1` | `91c2468b9bd8a411a918914794736c4d294529e3d6d4ea89c92f13422985f7b7` |
 | `v1.2` | `6ff159cb75a86051a46a705c13a4d9e22f26a093cdeefbf269c359d524cd9db9` |
+Foundry copies are named `pointwam_dexjoco_224_<tag>_<revision>`. The 256 x 256 DexJoCo frames in Foundry are
+separate renders, not versions of this dataset.
+## Download and check
 ```bash
+hf download chrockey/DexJoCo-preprocessed --repo-type dataset --revision v1.2 --local-dir DexJoCo-preprocessed
+hf download chrockey/DexJoCo-preprocessed versions/v1.2.sha256 --repo-type dataset --local-dir manifests
+cd DexJoCo-preprocessed && sha256sum -c --quiet ../manifests/versions/v1.2.sha256
 ```
+Add `--include "water_plant/*"` for one task. Always pin `--revision`. Folder links for training: PointWAM GitHub
+README, `DexJoCo training data`.
+## Contents
+| task | arms | demos | frames | cameras | GB |
+|---|---|---|---|---|---|
+| `bimanual_assembly` | 2 | 100 | 72,642 | back, handcam_rgb_left/right | 16.5 |
+| `bimanual_hanoi` | 2 | 100 | 109,105 | back, handcam_rgb_left/right | 26.3 |
+| `bimanual_microwave_cook` | 2 | 100 | 84,790 | back, handcam_rgb_left/right | 20.6 |
+| `bimanual_photograph` | 2 | 100 | 40,255 | back, handcam_rgb_left/right | 8.3 |
+| `bimanual_unlock_ipad` | 2 | 100 | 53,144 | back, handcam_rgb_left/right | 12.1 |
+| `click_mouse` | 1 | 100 | 39,982 | front, handcam_rgb | 5.4 |
+| `fold_glasses` | 1 | 100 | 57,900 | front, handcam_rgb | 8.8 |
+| `hammer_nail` | 1 | 100 | 25,034 | front, handcam_rgb | 4.3 |
+| `pick_bucket` | 1 | 100 | 46,910 | front, handcam_rgb | 7.3 |
+| `pinch_tongs` | 1 | 100 | 43,441 | front, handcam_rgb | 5.9 |
+| `water_plant` | 1 | 100 | 27,753 | front, handcam_rgb | 4.3 |
+Layout `<task>/<task>_demo_<k>_<date>_<time>.h5` (unlock_ipad files: `ipad_unlock_demo_*`). Counts are for `v1.2`.
+## H5 schema (T frames at 30 Hz)
+| key | shape | meaning |
+|---|---|---|
+| attrs | | `task`, `prompt`, `n_arm`, `n_frames`, `size`, `success`, `cameras`; `trim_start` in trimmed demos |
+| `actions` | (≤T, 23·n_arm) | per arm `[pos3, quat4, hand16]`, arms interleaved; may stop before T |
+| `proprio` | (T, 23·n_arm) | `[tcp pose 7·n_arm \| hand joints 16·n_arm]` (blocked) |
+| `qpos` | (T, nq) | MuJoCo `qpos` |
+| `body_names`, `is_robot`, `is_hand` | (B,) | bodies and robot / hand masks |
+| `xpos`, `xquat` | (T, B, 3/4) | body poses |
+| `<cam>/rgb`, `/depth`, `/body_id` | (T, 224, 224, ·) | RGB, depth (m, float16), body id |
+| `<cam>/intrinsic`, `/extrinsic` | (3, 3), (T, 4, 4) | camera intrinsics, per-frame pose |
+## How it was made
+- `brave-eai/dexjoco` `8d23b0f` + PointWAM's two patches; raw demos from `DexJoCo/DexJoCo-Datasets-Raw`.
+- `pointwam.preprocessing.dexjoco.replay_capture` (hold the last action 60 steps; hanoi 90, unlock_ipad 150,
+  photograph 0), then `extract_frames` at 224. `hammer_nail` from the `mocaptail` capture.
+- Every demo succeeds on replay. 83 needed recovery: 42 unlock_ipad (wrong iPad pose, brave-eai/dexjoco#27) and 41
+  marginal-contact demos (initial object pose search ≤ 3.2 mm). Lists: `experiments/dexjoco-demo-recovery.md` (GitHub).
+- `v1.2` trims 5 demos whose recording is frozen at the start (click_mouse 2, microwave_cook 3).
+- Checked against the raw recordings (`scripts/validate_dexjoco_frames.py`).